GPT, from scratch
A transformer language model assembled from the raw parts — tokenizer, attention, training loop — in PyTorch. The goal was to understand every line.
Mancera1/gpt-from-scratchJose Mancera / Machine Learning Engineer
I build systems that learn from data, at scale.
Machine-learning engineer at LinkedIn, working on video recommendations — vision-language models, retrieval, and content understanding.
01 About
Marine first.
Engineer after.
Before any of this, I was a sergeant in the Marine Corps. It taught me to ignore the noise and look for the truth from first principles — which, it turns out, is most of engineering.
I left, studied computer science, and started building machine-learning systems. The work I care about lives where research meets production — models that have to be genuinely good and actually ship to real people.
The ML is the fun part, not the whole point. What I like is hard problems, the people they're for, and doing the unglamorous work well.
02 Toolkit
What I work in
Less a list of tools, more the shape of the problems I like.
-
01
Recommendation & Retrieval
Ranking, embeddings, and the systems that decide what you see next.
-
02
Multimodal ML
Vision-language models and content understanding across video and text.
-
03
Models at Scale
Distributed training and serving large models without the whole thing falling over.
-
04
Home Base
Python and PyTorch, plus whatever the problem actually needs.
03 Path
How I got here
-
Now
LinkedIn
Machine-learning engineer on video recommendations, in Mountain View.
-
2023–2024
Amazon · SpaceX · AWS
A run of engineering internships across recommendation, infrastructure, and large-model serving.
-
School
Notre Dame · Baylor
An M.S. in progress at Notre Dame; a B.S. from Baylor before it.
-
Before
U.S. Marine Corps
Where it started. Out as a sergeant.
04 Selected Work
Things I've built
A few projects from outside the day job — built to learn, not to impress.
Retrieval-augmented generation
A RAG system that grounds a language model in a real document store and serves it on Kubernetes. Vector search in, sourced answers out.
Mancera1/rag-systemFully from-scratch RAG
The two above, wired together: my GPT is the generator inside my RAG — my tokenizer, my retrieval, my transformer, no external APIs — answering over a behavioural-data-science handbook in the demo below.
It's deliberately tiny and barely trained, and that's the point: it shows how the earliest GPTs actually began, and from these foundations you can see exactly where the later improvements come from.
Next steps aren't a better architecture — it's training on far more real data, on a GPU. I've only scratched the surface there, but the core idea is proven.